RocksDB

A highly resilient pattern involves writing the incoming order to a fast, local embedded database on the application server.

Something like RocksDB or an Squale write-ahead log.

Oh, because that disk IO operation takes microseconds.

Right.

The HTTP thread writes the order locally, immediately returns a 200 OK success code to the user, and goes back to the pool to serve the next customer.

The user's order is secured on local disk.

Then a completely isolated background thread running within the application reads from that local RocksDB and asynchronously pushes the data to Kafka.

Exactly.

Delete does not throw an error if the key that you're trying to delete does not exist.

And we found that, uh, many popular storage engines that were powering distributed systems, like for example, RocksDB, for instance, uh, had this interface.

In fact, even read-modify-writes in RocksDB, uh, are implemented as nil-externalizing operations where it's, uh, you can take, uh, read-modify-writes to a particular key, but it's not synchronously applied to the key and the results are not immediately returned.

It's only re- applied asynchronously in the storage engine, and the results are not returned at all synchronously to the user.

So then we-- that has got us thinking like, we can actually leverage this and then let's not delay durability.

And they weren't invented at Google, but Google really popularized them with LevelDB, which emerged out of the work on Bigtable and Spanner.

that got popularized into RocksDB.

I subsequently re-implemented one of these things, and this is what we use at Cockroach Labs.

It's called Pebble.

talking

pre-AI.

Pre-AI, right? This is when you're back doing this manually, right? You know, at some point, you know, we started out using a system called RocksDB, which is an LSM.

At some point, I think it's back in 2019, you ran into limitations with it.